⚡ electrical engineering

Powder-Dependent Temporal Observability Windows of In-Process Monitoring Signals in Laser Powder Bed Fusion

This study demonstrates that the temporal windows for predicting defects in Laser Powder Bed Fusion are not universal but depend on specific powder characteristics, necessitating tailored in-process monitoring strategies for different powder batches.

Original authors: Simon Sankare

Published 2026-10-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Simon Sankare

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of modern manufacturing, there is a technique called Laser Powder Bed Fusion that builds complex metal parts by melting tiny grains of metal powder with a high-powered laser. Imagine a machine that draws a shape in a bed of powder, melts it into a solid layer, and then repeats the process hundreds of times to create a three-dimensional object. This method is vital for making lightweight components for airplanes, medical implants, and energy systems. However, the process is notoriously difficult to control. Inside the machine, the laser creates a tiny, super-hot pool of molten metal that is invisible to the naked eye. As the machine builds upward, defects like tiny holes or cracks can form deep inside the part, hidden beneath layers that have already cooled. Because these flaws are buried, they often go unnoticed until the part is finished and cut open, leading to wasted time and expensive materials.

To solve this, engineers have developed "in-situ" monitoring systems. These are cameras and sensors mounted inside the machine that watch the molten metal pool as it forms, layer by layer. The hope has been that by watching the light and heat of this pool, the machine could instantly spot a problem and stop before a bad part is made. For a long time, the prevailing idea was that these sensors should work the same way for every type of metal powder and that a defect would show up on the sensor the moment it happened. But a new study suggests this view is too simple. The research reveals that the moment a sensor can actually tell if a part will be good or bad depends entirely on the specific type of powder being used, and that this "moment of truth" can happen at very different times during the build.

The researcher, led by Simon Sankare at Oerlikon AM Europe GmbH, set out to test this idea using a common aluminum alloy called AlSi10Mg. They did not just look at one type of powder; they gathered four distinct batches, each with different physical properties. One batch was made of highly round, smooth particles. Another had a mix of large and small particles. A third was made from recycled powder that had been used and melted down before, carrying more impurities. The fourth was a standard, single-sized powder. To ensure a fair test, they used the exact same machine settings and laser power for all four types, printing a total of 119 metal parts. They then examined the cross-sections of these parts to see which ones were solid and which ones had internal flaws.

The team then went back to the data recorded by the machine's sensors during the printing of each part. Instead of looking at the sensor data second-by-second or layer-by-layer, they looked for patterns that emerged over time. They asked a simple question: at what point in the building process does the sensor data start to reliably predict whether the final part will be defective? They scanned through the entire height of the builds, testing different time windows to see when the signals became useful.

The results were striking and showed that there is no single rule for all powders. For the powder with the highest roundness and smoothness, the sensors became very good at predicting quality very early in the process, specifically during the first 10 to 30 percent of the build height. This early window suggests that for this high-quality powder, the machine settles into a stable rhythm quickly, and any trouble shows up almost immediately. In contrast, the powder with a mix of particle sizes showed its best predictive signals in the middle of the build, around the 30 percent mark, once the process had moved past the initial heating phase.

The most surprising finding came from the recycled powder. For this batch, the sensors were almost useless for the first 70 percent of the build. The data only became informative in the final quarter of the process, between 70 and 75 percent of the height. This does not necessarily mean the defects only appeared at the end; rather, it suggests that the recycled powder is so variable that the machine needs to print a large amount of material before the sensor signals become clear enough to distinguish a good part from a bad one. The standard, single-sized powder was the only one that did not show a clear, reliable pattern at any specific time, indicating that for this material, the sensor signals were not a strong predictor of the final outcome.

The study also looked at how the data was analyzed. They found that looking at the average stability of the signal over time was more important than looking for single, extreme spikes. The most successful predictions came from measuring how much the signal fluctuated over a period, rather than just catching a single bad moment. Furthermore, the researchers tested whether these findings were just a result of the specific machine settings they used. They confirmed that the timing of these predictive windows was tied to the powder itself, not the laser settings. When they tried to use the "early warning" window from the smooth powder to predict the quality of the recycled powder, it failed completely. This proved that a monitoring strategy that works for one type of powder cannot be blindly copied to another.

This work changes how we think about monitoring these machines. It suggests that the idea of a universal "watchdog" that works the same way for every metal is flawed. Instead, the best way to monitor a build depends on the specific characteristics of the powder being used. For some materials, you need to watch the very beginning of the process; for others, you must wait until the part is nearly finished before the signals tell you anything useful. The study concludes that effective monitoring strategies should be adjusted to match the powder, rather than relying on a single, fixed rule for all situations. By understanding when the information becomes available, manufacturers can interpret their sensor data more accurately, potentially saving time and resources by knowing exactly when to trust the machine's eyes.

Technical Summary: Powder-Dependent Temporal Observability in Laser Powder Bed Fusion

Problem Statement
Final defects in Laser Powder Bed Fusion (LPBF) components often result from cumulative process evolution rather than isolated layer events. While in-process monitoring systems (e.g., melt pool sensors) are widely used, current interpretation frameworks typically assume that defects are instantaneous phenomena detectable within a single layer or that predictive information is uniformly distributed throughout the build. Furthermore, the influence of powder characteristics on the temporal localisation of predictive information remains largely unexplored. Most monitoring strategies rely on universal rules, potentially overlooking how different feedstocks (varying in chemistry, particle size, and morphology) alter the timing and nature of observable process instabilities.

Methodology
This study investigated the relationship between melt pool monitoring signal instability and part-level defect formation across four distinct AlSi10Mg powder batches:

Recycled feedstock: Unimodal, high oxygen/nitrogen/moisture content.
Standard unimodal: Unimodal, EN-AC-43000 chemistry.
High-sphericity: Unimodal, high sphericity.
Bimodal PSD: Bimodal particle size distribution.

Experimental Setup:

Platform: Colibrium-Concept Laser M2 system with a 1000 W Yb-fibre laser.
Design of Experiments (DOE): A common partial factorial design (4 levels of laser power, scanning speed, and trace spacing) was applied identically to all four powders to isolate powder effects from process parameter variations.
Dataset: 119 parts (approx. 30 per powder) were manufactured, each consisting of exactly 100 layers.
Monitoring: Data was extracted from the QMmeltpool 3D system, utilizing four metrics per layer: average intensity, intensity variance, average melt pool area, and area variance.
Quality Labeling: Parts were labeled "bad" or "good" based on a relative threshold (25th percentile of macro areal density) specific to each powder population to account for differing defect mechanisms.

Analytical Framework:

Temporal Aggregation: A two-dimensional window scan was performed over the build height (normalized 0–100%). Windows were defined by a start position ( $s$ ) and width ( $w$ ).
Aggregation Statistics: Three statistics were computed within each window: Standard Deviation ( $A_{std}$ ), Maximum ( $A_{max}$ ), and 95th Percentile ( $A_{p95}$ ). Mean aggregation was excluded to preserve sensitivity to extreme events.
Validation: Predictive performance was quantified using ROC-AUC. To ensure robustness and avoid overfitting, the study employed:
- Leave-One-Part-Out Cross-Validation (LOPO-CV): The optimal monitoring window was re-selected within each training fold.
- Residualisation: Signals were adjusted for nominal process parameters (power, speed, hatch) to verify they contained information beyond commanded settings.
- Robustness Checks: Layer shuffle tests, cross-powder transfer tests, and label sensitivity analyses were conducted.

Key Results
The study identified that predictive information is neither instantaneous nor universal; it is confined to specific, powder-dependent temporal windows:

High-Sphericity Powder: The strongest predictive signature occurred during the initial thermal transient (10–30% build height). The optimal metric was the variance of the melt pool area.
Bimodal PSD Powder: Predictive information emerged during the early steady-state regime (centered around 30% build height), primarily associated with the temporal variability of the photodiode intensity signal.
Recycled Feedstock: A robust predictive window appeared only in the late stage of the build (70–75% height), utilizing the standard deviation of the average melt pool area.
Standard Unimodal Powder: No statistically robust temporal monitoring signature was identified. Its apparent predictive capability was largely attributed to correlations with nominal process parameters rather than independent process-state signals.

Aggregation and Robustness:

Aggregation: Temporal standard deviation consistently provided the highest predictive performance, suggesting that sustained fluctuations are more informative than isolated peak events.
Transferability: Monitoring windows were not transferable between powders. Applying a window optimized for one powder to another often resulted in AUC values near or below random chance (0.5), and in some cases, the signals became anti-correlated with quality.
Independence: For the recycled, bimodal, and high-sphericity powders, predictive performance remained robust even after residualising for nominal process parameters, confirming the signals captured intrinsic process-state information.

Significance and Claims
The paper concludes that the temporal localisation of predictive information is a powder-dependent process characteristic.

Observability Horizon: The "observability horizon"—the amount of process history required before quality differences become distinguishable—varies significantly based on powder properties (e.g., thermal conductivity, flowability, and chemical state).
Strategy Adjustment: Effective LPBF monitoring strategies should be adjusted to specific powder characteristics rather than relying on universal, machine-wide rules. A strategy qualified for one feedstock cannot be assumed valid for another, even under identical machine settings.
Methodological Contribution: The proposed framework serves as an analytical tool to identify when monitoring signals become informative. It suggests that the temporal position of the monitoring window itself could serve as a descriptor for powder qualification or reuse assessment.

The author emphasizes that these findings are hypothesis-generating, based on a specific set of four powders and a single machine platform, and call for further investigation to isolate the specific physical contributions of particle morphology, size distribution, and oxidation state.

Technical Summary: Powder-Dependent Temporal Observability in Laser Powder Bed Fusion

More like this